Source is preserved and inspectable.
The private collection contains 133 files across eight project groups. Its prepared Git tree matched the uploaded GitHub tree. All 23 Python files passed syntax parsing; this is not an end-to-end test result.
● Experiments by Anton / 59notna
Less managing.
More making.
A small lab for putting agents to work.
Two projects exploring how to turn a rough idea into a plan, a build, and something you can actually check.
Independent projects built with Codex and OpenAI tools.
WORKFLOW / ASTRA + FLASHExplorable prototype
THE STARTING POINT“Build the thing.
Show that it works.”
Show that it works.”
⌘
ASTRAPlan & decide
01↗
FLASHBuild & test
02✓
ASTRAReview the evidence
03SPEC → PATCH → EVIDENCE
The workflow
Give each model
a clear job.
Astra Flash Orchestrator packages a division of work: planning and acceptance stay with Astra; implementation and routine verification go to Flash. Explore what changes hands at each stage.
ASTRA / PLANNING
Make the assignment concrete.
Define the objective, boundaries, acceptance checks, and files the worker owns. A useful handoff gives the builder enough context to execute a complete slice of work.
objective: Add an invoice status filter scope: UI + query + focused tests acceptance: Preserve existing sorting handoff: Spec, constraints, checks
✓ Reversible installer✓ Plan validation✓ Structured review
This explains the packaged workflow. It does not run agents or make provider requests on this page.
Launch Desk / interactive demo
From a brief
to a next step.
Edit a launch brief. Run the actual deterministic tools used by Launch Desk to create tasks, group owners, and identify missing details. Everything here runs in your browser.
The working plan
Ready to explore—/100
Brief readiness
Keyword-based coverage score.
Not a release approval.
Behind the full app
OpenAI Agents SDK · React · Express · streamed planning responses
The evidence
Show the work.
State the limits.
Working code matters. So does being precise about what has actually been checked.
Download the project brief ↓Orchestrator includes offline validation.
The package records 63 passing offline tests on macOS for an earlier candidate. That report covers installation and configuration safeguards. It does not establish a fresh Windows result or successful live routing for every provider.
A fair efficiency comparison.
An earlier local benchmark compared different task mixes and estimated costs. It is useful exploratory evidence, but not a controlled quality comparison or a verified billing reduction. A repeatable, same-task evaluation is the next step.
Built by 59notna
Small experiments.
Useful things.
This showcase brings two projects out of the local workspace and into a form you can explore. The wider collection also includes media tools, Remotion compositions, a lead pipeline, and Windows scripts.
Visit my GitHub ↗